Fixes the pgstats side of BUG #19520, where a stale concurrent function/procedure call can create and drop a function stats entry for an already-dropped procedure OID while the DROP transaction is still in post-commit stats cleanup.

Before this change, the second drop of the same shared stats entry raised:

ERROR: trying to drop stats entry already dropped: kind=function ...

When that ERROR happened after RecordTransactionCommit(), xact abort handling escalated to:

PANIC: cannot abort transaction ..., it was already committed

This patch makes duplicate drops of an already-dropped shared stats entry behave like "not freed yet" rather than ERRORing, so normal pgstat entry-ref GC can finish the cleanup. The stale function-call path also requests that GC when its stats drop cannot free immediately.

The test adds injection points to deterministically schedule the race:

• stale CALL proc_test() resolves the procedure OID and waits before function stats init
• concurrent DROP PROCEDURE proc_test() waits in post-commit function stats cleanup
• stale CALL creates/drops function stats for the removed OID while another session holds a local stats ref
• DROP cleanup attempts the duplicate stats drop

Validation:

• Temporarily restoring the old elog(ERROR) reproduces the original ERROR -> COMMIT-state abort warning -> PANIC path in the new isolation spec.
• meson setup build -Dinjection_points=true -Dcassert=true
• ninja -C build
• meson test -C build --suite setup --print-errorlogs
• env PG_TEST_TIMEOUT_DEFAULT=60 meson test -C build 'injection_points/isolation' 'injection_points/regress' --print-errorlogs